News · 2026-08-23

SenseNova generates images in pixels with no VAE in the middle

SenseTime has released SenseNova-U1.5-8B-MoT under an Apache 2.0 license, an open-weight model that both understands and generates images without the two components almost every comparable system depends on: a separate vision encoder for reading images, and a variational autoencoder for compressing them during generation. It works on near-lossless pixels directly. The architecture is the news; the image quality is the argument for it.

Key facts

• Roughly 18 billion parameters and about 50 GB of weights, despite the "8B" in the name.
• Released under Apache 2.0 by SenseTime, with the umbrella paper at arXiv 2605.12500.
• The architecture, called NEO-unify, uses one mixture-of-transformers backbone: cross-entropy training for text, pixel flow matching for vision.
• Primary source: the Hugging Face model card, with code at GitHub.

To see why this is a real architectural bet rather than a tweak, it helps to know what the standard pipeline does. A modern image model does not generate pixels. It generates a compressed representation -- typically eight times smaller in each dimension -- produced by a variational autoencoder trained separately, and then decodes that representation back into an image. The compression is what makes generation affordable: denoising a 64-by-64 latent grid is vastly cheaper than denoising a 512-by-512 image. The cost is that the generator never sees a real pixel during training, and everything the compressor discarded -- fine text, small faces, high-frequency texture -- is gone before generation starts. Reading images works the same way in reverse, through a frozen vision encoder that produces a summary the language model consumes.

SenseNova's NEO-unify removes both. One backbone handles text and vision, trained with autoregressive cross-entropy on the text side and flow matching on the pixel side. The mixture-of-transformers arrangement means the two modalities have their own parameters within a shared attention structure rather than sharing everything or living in separate models.

The engineering problem this creates is obvious: producing full-resolution pixels directly is expensive. U1.5's answer is a change to the image head, replacing patchwise RGB prediction with progressive patch-joint reconstruction. Visual tokens are reshaped into a two-dimensional grid and upsampled through several pixel-shuffle stages, with small convolutions between them -- essentially building the picture up in resolution steps rather than predicting every pixel independently. That decoder is in the public source, which is the strongest evidence that the "no VAE" claim describes real code and not marketing.

The analogy is the difference between a translator working from a summary and one working from the original text. The summary is faster to handle and usually sufficient. But you cannot recover a detail the summarizer decided not to include, no matter how good your translation is.

It is also a full editor, not just a text-to-image model. The card documents natural-language edits, single- and multi-image references, insertion and replacement, and region control through bounding boxes and drawn markers. SenseTime's published comparison tables put it alongside closed systems including GPT Image 2 and Qwen Image 2 -- clearly ahead of the previous U1 generation, and competitive rather than dominant against the strongest closed editors.

There is a second reason the release is being watched by practitioners rather than only researchers. Because there is no separate autoencoder, there is no separate autoencoder to swap, fine-tune, or blame -- the whole image path is one set of weights under one permissive license. That matters for downstream tooling: local runtimes moved on it within days, with feature requests opened against ComfyUI and stable-diffusion.cpp to add support, which is the practical measure of whether an open image model gets used or admired. The repository also documents a 24 GB-class fast mode and a lower-memory quantized path, so the barrier is a large consumer card rather than a datacenter node -- unlike most models at this parameter count.

Why it matters: the encoder-plus-latent-autoencoder stack has been so standard for so long that its costs are rarely questioned. An Apache-2.0 model that removes both and stays competitive is the first serious open evidence that the standard stack is a convenience rather than a requirement -- and convenience assumptions are exactly the ones that get overturned when compute gets cheaper.

The caveats are practical. The name undersells the size: this is an 18-billion-parameter model at roughly 50 GB of weights, not an 8-billion-parameter one, and running it comfortably means quantized or offloaded execution on consumer hardware. The documentation describes native 4K generation, but the public inference examples still default to 2048 by 2048, so the 4K path is only partly spelled out. And while the umbrella U1 paper is public, there is no separate technical report for U1.5 yet, so the specific changes in this version are documented by repository notes rather than a reviewed paper.

Primary source, verified: read the paper → (arXiv 2605.12500)

Ground Truth. (2026, August 23). SenseNova generates images in pixels with no VAE in the middle. Ground Truth. https://groundtruth.day/news/sensenova-generates-images-in-pixels-with-no-vae-in-the-middle.html

@misc{groundtruth:sensenova-generates-images-in-pixels-with-no-vae-in-the-middle,
  title  = {SenseNova generates images in pixels with no VAE in the middle},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/sensenova-generates-images-in-pixels-with-no-vae-in-the-middle.html}
}

Topics: image-generation · open-weights · multimodal · architecture · image-editing